ECE 1747H : Parallel Programming

1. ECE 1747H : Parallel Programming Message Passing (MPI)

2. Explicit Parallelism • Same thing as multithreading for shared memory. • Explicit parallelism is more common with message passing. • User has explicit control over processes. • Good: control can be used to performance benefit. • Bad: user has to deal with it.

3. Distributed Memory - Message Passing mem1 mem2 mem3 memN proc1 proc2 proc3 procN network

4. Distributed Memory - Message Passing • A variable x, a pointer p, or an array a[] refer to different memory locations, depending of the processor. • In this course, we discuss message passing as a programming model (can be on any hardware)

5. What does the user have to do? • This is what we said for shared memory: • Decide how to decompose the computation into parallel parts. • Create (and destroy) processes to support that decomposition. • Add synchronization to make sure dependences are covered. • Is the same true for message passing?

6. Another Look at SOR Example for some number of timesteps/iterations { for (i=0; i<n; i++ ) for( j=0; j<n, j++ ) temp[i][j] = 0.25 * ( grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1] ); for( i=0; i<n; i++ ) for( j=0; j<n; j++ ) grid[i][j] = temp[i][j]; }

7. Shared Memory grid temp 1 1 2 2 3 3 4 4 proc1 proc2 proc3 procN

8. Message-Passing Data Distribution (only middle processes) grid grid 2 3 temp temp 2 3 proc2 proc3

9. Is this going to work? Same code as we used for shared memory for( i=from; i<to; i++ ) for( j=0; j<n; j++ ) temp[i][j] = 0.25*( grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1]); No, we need extra boundary elements for grid.

10. Data Distribution (only middle processes) grid grid 2 3 temp temp 2 3 proc2 proc3

11. Is this going to work? Same code as we used for shared memory for( i=from; i<to; i++) for( j=0; j<n; j++ ) temp[i][j] = 0.25*( grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1]); No, on the next iteration we need boundary elements from our neighbors.

12. Data Communication (only middle processes) grid grid proc2 proc3

13. Is this now going to work? Same code as we used for shared memory for( i=from; i<to; i++ ) for( j=0; j<n; j++ ) temp[i][j] = 0.25*( grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1]); No, we need to translate the indices.

14. Index Translation for( i=0; i<n/p; i++) for( j=0; j<n; j++ ) temp[i][j] = 0.25*( grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1]); Remember, all variables are local.

15. Index Translation is Optional • Allocate the full arrays on each processor. • Leave indices alone. • Higher memory use. • Sometimes necessary (see later).

16. What does the user need to do? • Divide up program in parallel parts. • Create and destroy processes to do above. • Partition and distribute the data. • Communicate data at the right time. • (Sometimes) perform index translation. • Still need to do synchronization? • Sometimes, but many times goes hand in hand with data communication.

17. Message Passing Systems • Provide process creation and destruction. • Provide message passing facilities (send and receive, in various flavors) to distribute and communicate data. • Provide additional synchronization facilities.

18. MPI (Message Passing Interface) • Is the de facto message passing standard. • Available on virtually all platforms. • Grew out of an earlier message passing system, PVM, now outdated.

19. MPI Process Creation/Destruction MPI_Init( int argc, char **argv ) Initiates a computation. MPI_Finalize() Terminates a computation.

20. MPI Process Identification MPI_Comm_size( comm, &size ) Determines the number of processes. MPI_Comm_rank( comm, &pid ) Pid is the process identifier of the caller.

21. MPI Basic Send MPI_Send(buf, count, datatype, dest, tag, comm) buf: address of send buffer count: number of elements datatype: data type of send buffer elements dest: process id of destination process tag: message tag (ignore for now) comm: communicator (ignore for now)

22. MPI Basic Receive MPI_Recv(buf, count, datatype, source, tag, comm, &status) buf: address of receive buffer count: size of receive buffer in elements datatype: data type of receive buffer elements source: source process id or MPI_ANY_SOURCE tag and comm: ignore for now status: status object

23. MPI Matrix Multiply (w/o Index Translation) main(int argc, char *argv[]) { MPI_Init (&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &myrank); MPI_Comm_size(MPI_COMM_WORLD, &p); from = (myrank * n)/p; to = ((myrank+1) * n)/p; /* Data distribution */ ... /* Computation */ ... /* Result gathering */ ... MPI_Finalize(); }

24. MPI Matrix Multiply (w/o Index Translation) /* Data distribution */ if( myrank != 0 ) { MPI_Recv( &a[from], n*n/p, MPI_INT, 0, tag, MPI_COMM_WORLD, &status ); MPI_Recv( &b, n*n, MPI_INT, 0, tag, MPI_COMM_WORLD, &status ); } else { for( i=1; i<p; i++ ) { MPI_Send( &a[from], n*n/p, MPI_INT, i, tag, MPI_COMM_WORLD ); MPI_Send( &b, n*n, MPI_INT, I, tag, MPI_COMM_WORLD ); } }

25. MPI Matrix Multiply (w/o Index Translation) /* Computation */ for ( i=from; i<to; i++) for (j=0; j<n; j++) { C[i][j]=0; for (k=0; k<n; k++) C[i][j] += A[i][k]*B[k][j]; }

26. MPI Matrix Multiply (w/o Index Translation) /* Result gathering */ if (myrank!=0) MPI_Send( &c[from], n*n/p, MPI_INT, 0, tag, MPI_COMM_WORLD); else for (i=1; i<p; i++) MPI_Recv( &c[from], n*n/p, MPI_INT, i, tag, MPI_COMM_WORLD, &status);

27. MPI Matrix Multiply (with Index Translation) main(int argc, char *argv[]) { MPI_Init (&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &myrank); MPI_Comm_size(MPI_COMM_WORLD, &p); from = (myrank * n)/p; to = ((myrank+1) * n)/p; /* Data distribution */ ... /* Computation */ ... /* Result gathering */ ... MPI_Finalize(); }

28. MPI Matrix Multiply (with Index Translation) /* Data distribution */ if( myrank != 0 ) { MPI_Recv( &a, n*n/p, MPI_INT, 0, tag, MPI_COMM_WORLD, &status ); MPI_Recv( &b, n*n, MPI_INT, 0, tag, MPI_COMM_WORLD, &status ); } else { for( i=1; i<p; i++ ) { MPI_Send( &a[from], n*n/p, MPI_INT, i, tag, MPI_COMM_WORLD ); MPI_Send( &b, n*n, MPI_INT, I, tag, MPI_COMM_WORLD ); } }

29. MPI Matrix Multiply (with Index Translation) /* Computation */ for ( i=0; i<n/p; i++) for (j=0; j<n; j++) { C[i][j]=0; for (k=0; k<n; k++) C[i][j] += A[i][k]*B[k][j]; }

30. MPI Matrix Multiply (with Index Translation) /* Result gathering */ if (myrank!=0) MPI_Send( &c, n*n/p, MPI_INT, 0, tag, MPI_COMM_WORLD); else for( i=1; i<p; i++ ) MPI_Recv( &c[from], n*n/p, MPI_INT, i, tag, MPI_COMM_WORLD, &status);

31. Running a MPI Program • mpirun <program_name> <arguments> • Interacts with a daemon process on the hosts. • Causes a Unix process to be run on each of the hosts. • May only run in interactive mode (batch mode may be blocked by ssh)

32. ECE1747 Parallel Programming Message Passing (MPI) Global Operations

33. What does the user need to do? • Divide program in parallel parts. • Create and destroy processes to do above. • Partition and distribute the data. • Communicate data at the right time. • (Sometimes) perform index translation. • Still need to do synchronization? • Sometimes, but many times goes hand in hand with data communication.

34. MPI Process Creation/Destruction MPI_Init( int *argc, char **argv ) Initiates a computation. MPI_Finalize() Finalizes a computation.

35. MPI Process Identification MPI_Comm_size( comm, &size ) Determines the number of processes. MPI_Comm_rank( comm, &pid ) Pid is the process identifier of the caller.

36. MPI Basic Send MPI_Send(buf, count, datatype, dest, tag, comm) buf: address of send buffer count: number of elements datatype: data type of send buffer elements dest: process id of destination process tag: message tag (ignore for now) comm: communicator (ignore for now)

37. MPI Basic Receive MPI_Recv(buf, count, datatype, source, tag, comm, &status) buf: address of receive buffer count: size of receive buffer in elements datatype: data type of receive buffer elements source: source process id or MPI_ANY_SOURCE tag and comm: ignore for now status: status object

38. Global Operations (1 of 2) • So far, we have only looked at point-to-point or one-to-one message passing facilities. • Often, it is useful to have one-to-many or many-to-one message communication. • This is what MPI’s global operations do.

39. Global Operations (2 of 2) • MPI_Barrier • MPI_Bcast • MPI_Gather • MPI_Scatter • MPI_Reduce • MPI_Allreduce

40. Barrier MPI_Barrier(comm) Global barrier synchronization, as before: all processes wait until all have arrived.

41. Broadcast MPI_Bcast(inbuf, incnt, intype, root, comm) inbuf:address of input buffer (on root); address of output buffer (elsewhere) incnt: number of elements intype: type of elements root: process id of root process

42. Before Broadcast inbuf proc0 proc1 proc2 proc3 root

43. After Broadcast inbuf proc0 proc1 proc2 proc3 root

44. Scatter MPI_Scatter(inbuf, incnt, intype, outbuf, outcnt, outtype, root, comm) inbuf: address of input buffer incnt: number of input elements intype: type of input elements outbuf: address of output buffer outcnt: number of output elements outtype: type of output elements root: process id of root process

45. Before Scatter inbuf outbuf proc0 proc1 proc2 proc3 root

46. After Scatter inbuf outbuf proc0 proc1 proc2 proc3 root

47. Gather MPI_Gather(inbuf, incnt, intype, outbuf, outcnt, outtype, root, comm) inbuf: address of input buffer incnt: number of input elements intype: type of input elements outbuf: address of output buffer outcnt: number of output elements outtype: type of output elements root: process id of root process

48. Before Gather inbuf outbuf proc0 proc1 proc2 proc3 root

49. After Gather inbuf outbuf proc0 proc1 proc2 proc3 root

50. Broadcast/Scatter/Gather • Funny thing: these three primitives are sends and receives at the same time (a little confusing sometimes). • Perhaps un-intended consequence: requires global agreement on layout of array.